Reviewer #3: The paper introduces a novel Unit-Merge Network that stores binary patterns as integer units and retrieves memories through a hierarchical compound-key cohesion network. However, some points remain insufficiently clarified or were only partially addressed. In particular, the inclusion of a dedicated ablation study section is still missing and was not incorporated into the revised manuscript. Incorporating an ablation study is essential to showcase the contribution of each component of the proposed system. Kindly reconsider the following points, as they are only partially addressed.

The author was requested to provide mathematical formalisation, as this comment was only partially addressed. It was also suggested to include an ablation section to showcase the contribution of each component of the proposed system. In the revised manuscript, the author states that Sections 4.1 and the Conclusions were updated to address the ablation request. However, these sections do not contain a systematic ablation experiment. Although the manuscript provides some observations concerning unit size, this does not establish the individual contribution of the main components of the proposed architecture. It is recommended to provide a compact ablation study examining, where applicable, unit size, cohesion-network design, compound-key generation, and retrieval strategy. Quantitative results should be reported to demonstrate how each design choice affects retrieval accuracy, robustness and/or computational cost. Again, while checking the Mathematical formalisation. The newly added discussion in Section 4.1 provides an information-theoretic interpretation of the proposed network and discusses the empirical behaviour. However, it does not provide a formal mathematical definition of the compound-key operation, a rigorous mathematical description of the Unit-Merge construction, or a formal proof/argument establishing retrieval correctness. It is recommended that the authors introduce precise mathematical notation for the unit and compound-key operations, formally describe the cohesion-network construction, and provide a proof or clearly stated conditions under which a stored pattern can be uniquely retrieved. The O(log n) complexity claim should also be supported by a rigorous derivation, or explicitly presented as an empirical observation rather than a theoretical bound.

In my previous comment, I asked to expand the SOTA comparison. The author addressed this issue by expanding the related-work section by mentioning Modern Hopfield Networks and related associative-memory/Transformer-based models. However, the experimental evaluation still does not provide a controlled comparison with recent associative-memory methods. The comparison in Section 5.4 mainly reports results from different studies using different datasets, tasks and experimental protocols, and therefore does not constitute a normalized SOTA comparison. It is recommended either to include a controlled comparison with appropriate recent associative-memory baselines or, if this is outside the intended scope, to explicitly acknowledge the limitation and moderate claims such as "state-of-the-art" and "competitive with other modern architectures."

It is suggested to expand the literature review to at least 30 peer-reviewed references from top-tier journals and conferences.


Reply
-----
I have made the empirical evaluation more clear, shown by the highlighted text. While I do not have a formal mathematical definition, I have made a comparison with the Information Theory mathematical definition and found the complexity to be similar. I think that this also covers the ablation aspect, where there is a binary search factored by a number of units. I do not really want to add more here.

The papers quoted are state of the art and there are several results listed. I have not implemented them myself but retrieved the information from other papers, but I think that it would justify the quantitative comparisons.

Transformers is the first paragraph of the related work section. I have added another reference, but I am not at all convinced that the architecture is suitable because it is much too complicated for this paper.